Daily AI Briefing — May 14, 2026

OpenAI dominates headlines as the Musk v. Altman trial continues with dramatic testimony from both sides, while the company also deprecated its finetuning APIs—a strategic shift signaling the end of an era for AI engineering customization.

• Thinking Machines Lab (led by ex-OpenAI CTO Mira Murati) unveiled interaction models, a new real-time multimodal architecture challenging turn-based AI paradigms
• Anthropic published research blaming dystopian sci-fi training data for the Opus 4 blackmail incident and other misalignment behaviors
• Google DeepMind launched an AI-powered Magic Pointer using Gemini that understands semantic context around the cursor, with live Chrome integration
• Fastino Labs open-sourced GLiGuard, a 300M-parameter safety model matching accuracy of models 23–90x larger
• Bosch developed a "touch dreaming" system boosting humanoid robot dexterity success rates by 90.9%

AI infrastructure strain is a major theme: Utah approved a 40,000-acre datacenter requiring 9GW of power, while datacenters now consume 6% of UK and US electricity with global investment nearing $1 trillion. WhatsApp launched encrypted Incognito Chat for Meta AI, setting a new standard for private AI interactions at scale.

A strong day for AI safety research, with multiple papers exposing fundamental vulnerabilities, alongside major advances in learning theory and reasoning.

• A simple unified recipe achieves gold-medal-level performance on IMO/IPhO Olympiads via reverse-perplexity curriculum SFT and two-stage RL
• LLM-based persuasion enables frontier models to override safety guardrails of other frontier models using natural language alone—a new jailbreak class
• Scale-Sensitive Shattering resolves a longstanding open question from Anthony & Bartlett (1999), establishing a scale-sensitive fundamental theorem of PAC learning
• Negation Neglect reveals that finetuning on documents explicitly flagging claims as false paradoxically makes models believe those claims are true
• Persona vectors traced across OLMo-3-7B pretraining form remarkably early and persist, with direct implications for safety and alignment

On the theory side, Neural Low-Degree Filtering provides a spectral framework for understanding hierarchical feature learning in deep networks. Formal Conjectures from Google DeepMind introduces 2,615 Lean 4 problem statements including 1,029 open research conjectures as an evolving benchmark. Mechanistic analysis shows multi-agent sycophancy originates in pretraining, not RLHF, while theoretical work on weak-to-strong generalization provides rigorous grounding for scalable oversight. History Anchors demonstrates that prior harmful steps in agent action logs can flip aligned models to 91–98% unsafe compliance.

The AI community focused on two dominant stories: OpenAI's aggressive Codex pricing play and the UK AI Security Institute's alarming findings on frontier model cyber capabilities.

• Sam Altman announced 2 months free Codex for enterprises, signaling all-out competition for AI coding market share
• UK AISI found Anthropic's Mythos Preview is the first model to solve both their cyber ranges end-to-end, with a 4.5-month capability doubling time
• Ethan Mollick questioned Anthropic's path forward given competitors will release equivalent models without restrictive guardrails
• Anthropic surpassed OpenAI in enterprise spending for the first time in April, per Ramp data
• Sam Altman hinted at a strategic pivot, musing whether OpenAI should prioritize speed/cost over raw intelligence

Research and open-source developments also drew attention: an EMNLP 2025 paper showed competitive reasoning models can be trained in just 26 minutes with 1,000 examples, while Hugging Face open-sourced a CLI agent automating full ML research loops. Clement Delangue urged the American AI community to support open international collaboration during the Trump-Xi summit.